Comparison of Visual Place Recognition Methods for UAV Imagery

In many real world applications (natural disaster management, urban development, infrastructure inspection) Unmanned Aerial Vehicles (UAVs) perform flights on different times, for scene image acquisition. Visual Place Recognition (VPR) methods can match newly acquired images with older ones, when the new and/or the old ones are not georeferenced. Most VPR solutions are based on image retrieval, where a query image scene is visually compared with that of many related images in a database, and the most relevant ones are retrieved. Deep learning-based VPR performance relies a lot on image dataset acquisition conditions, e.g. structured/unstructured scene visualization, single/multi-view image acquisition, illumination variations, or on-road/aerial view. Most of VPR methods are trained and tested on on-road views. This paper addresses the issue of image retrieval performance when large image databases are employed. To this end, we perform a comparison of some state of the art VPR methods on UAV image datasets, where the amount of database images is scaled, examine how well they generalize, and expand on some dataset creation gaps for this task.